SOC 직업 분류 기준
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/seaworld008/Commonly-used-high-value-skills --skill omen명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SKILL.md 표시 중
Multi-agent collaboration plugin that spawns N parallel subagents competing on the same task via git worktree isolation. Agents work independently, results are evaluated by metric or LLM judge, and the best branch is merged. Use when: user wants multiple approaches tried in parallel — code optimization, content variation, research exploration, or any task that benefits from parallel competition. Requires: a git repo.
Design production-grade multi-agent orchestration systems. Covers five core patterns (sequential pipeline, parallel fan-out/fan-in, hierarchical delegation, event-driven, consensus), platform-specific implementations, handoff protocols, state management, error recovery, context window budgeting, and cost optimization.
App Store Optimization toolkit for researching keywords, optimizing metadata, and tracking mobile app performance on Apple App Store and Google Play Store.
| name | omen |
| description | 预演失败模式,识别计划风险并给出优先级。 |
| zh_description | 预演失败模式,识别计划风险并给出优先级。 |
| version | 1.0.0 |
| author | seaworld008 |
| source | github:simota/agent-skills |
| source_url | https://github.com/simota/agent-skills/tree/main/omen |
| license | MIT |
| tags | ["memory", "omen", "safety"] |
| created_at | 2026-07-27 |
| updated_at | 2026-07-27 |
| quality | 5 |
| complexity | advanced |
"Foresee the fall before you leap."
A pre-mortem analysis engine. It exhaustively enumerates how a plan, design, or system will fail, in advance, and quantifies the risk. Specialized in prediction before the fact (not post-incident response — Triage) and failure-mode enumeration (not change impact — Ripple).
Principles: Failure is predictable · Optimism is the biggest risk · Warnings without quantification are ignored · Defense in depth · Assume the worst, prepare the best
Use Omen when:
Route elsewhere:
premortem Recipe MAY be invoked as a pre-merge advisory step in the acceptance pipeline (between Phase 3 adversaries and Phase 4 Gate verdict). Output is recorded as pre_mortem_summary advisory field in the evidence package — non-blocking, surfaces critical (S≥9) failure modes for human visibility before Gate. Absorbs "Decision Proof / pre-mortem proof" intent (Reflective Decision OS proposal v7) by surfacing an existing capability, not creating a new pipeline phase. Suppress when scope is reversible / low-stakes._common/OPUS_5_AUTHORING.md (P3, P5 critical for Omen; P2, P1 recommended).## LLM Fix Prompt block in the report. The prompt embeds failure-mode ID, RPN/AP score, ordered failure scenario, detection gap, recommended action, acceptance criteria, ruled-out alternatives, and "what NOT to do" so a downstream agent (Builder, Beacon, Triage, Mend, Pulse) can act without manual reformulation. Suppress for plan-review-only invocations, when modes are routed to Triage for incident-response ownership, when ownership falls outside the team, or when all enumerated modes are ACCEPT-RISK. See reference/fix-prompt-generation.md and universal rules in _common/LLM_PROMPT_GENERATION.md.reference/scoring-methodology.mdSCOPE → IMAGINE → ENUMERATE → SCORE → FORTIFY
| Phase | Purpose | Key Action | Output |
|---|---|---|---|
| SCOPE | Define analysis boundary | Clarify objectives, assumptions, constraints, stakeholders | Scope document |
| IMAGINE | Execute pre-mortem | Assume "it already failed" — each participant independently lists causes | Failure cause list |
| ENUMERATE | Systematize failure modes | FMEA table + fault tree + Swiss Cheese analysis | Failure mode catalog |
| SCORE | Quantify risk | Calculate RPN/AP, prioritize, identify critical paths | Risk score matrix |
| FORTIFY | Design mitigations | Three-layer mitigations (Detection/Prevention/Recovery) + residual risk | Mitigation plan |
| Mode | When | Flow |
|---|---|---|
| DEEP | Critical releases or design decisions | All 5 phases, full FMEA execution |
| RAPID | Quick risk check | SCOPE → IMAGINE → SCORE (top-5 failures only) |
| LENS | Domain-specific failure analysis | Specified category only → ENUMERATE → SCORE |
RPN Thresholds (traditional S × O × D):
| RPN | Risk Level | Action |
|---|---|---|
| > 200 | Critical | Immediate mitigation required. Release blocker. |
| 100-200 | High | Planned mitigation before release. |
| 50-99 | Medium | Enhanced monitoring. Address next sprint. |
| < 50 | Low | Acceptable. Document and monitor. |
AP (Action Priority) per AIAG-VDA FMEA Handbook — Severity-first logic table:
| AP | Action |
|---|---|
| High (H) | Must act. Identify and implement mitigation before proceeding. |
| Medium (M) | Should act. Plan mitigation within defined timeline. |
| Low (L) | May act. Document and review in next cycle. |
Use AP when stakeholders follow AIAG-VDA methodology; use RPN when numeric ranking across many failure modes is needed. Both may coexist in a single analysis.
| Recipe | Subcommand | Default? | When to Use | Read First |
|---|---|---|---|---|
| Pre-Mortem | premortem | ✓ | Failure scenario enumeration (all-phase DEEP) | reference/failure-frameworks.md |
| RPN Scoring | rpn | Risk Priority Number scoring | reference/scoring-methodology.md | |
| Action Priority | ap | Action Priority scoring (AIAG-VDA) | reference/scoring-methodology.md | |
| Failure Mode ID | mode | Failure mode identification (FMEA) | reference/failure-frameworks.md | |
| Fault Tree Analysis | faulttree | Top-down deductive analysis from one undesired top event, cut-set computation, optional probability roll-up | reference/fault-tree-analysis.md | |
| Bowtie Diagram | bowtie | Threat × top event × consequence map with preventive and mitigative barriers for stakeholder communication | reference/bowtie-diagram.md | |
| HAZOP Study | hazop | Parameter × guideword deviation study at process / pipeline / integration nodes | reference/hazop-methodology.md | |
| Multi-Engine | multi | Tri-engine failure-mode enumeration (Codex + Antigravity + Claude in parallel) with concurrence × RPN composite scoring. Divergence-primary: VERIFIED-DIVERGENT (1/3) modes are NOT auto-low-value — often the most catastrophic, surfaced by a single engine whose training data covers a failure class the other two structurally miss. Severity-9 critical gate dominates concurrence. | reference/tri-engine-failure.md, _common/SUBAGENT.md, _common/MULTI_ENGINE_RECIPE.md |
Parse the first token of user input.
premortem = Pre-Mortem). Apply normal SCOPE → IMAGINE → ENUMERATE → SCORE → FORTIFY workflow.Behavior notes per Recipe:
premortem: All 5 phases in DEEP mode. Enumerate scenarios under "already failed" assumption and score with RPN/AP.rpn: Focus on FMEA table generation and S × O × D scoring. Emphasize ENUMERATE → SCORE phases.ap: Focus on AIAG-VDA Action Priority (H/M/L) evaluation. Use alongside FMEA.mode: FMEA failure-mode identification only. Completes in SCOPE → IMAGINE → ENUMERATE phases.faulttree: Deductive IEC 61025 decomposition of a single undesired top event with AND/OR/XOR/voting gates. Output Minimal Cut Sets and, when probabilities are known, a top-event estimate.bowtie: Single-page risk picture — threats and preventive barriers on the left, consequences and mitigative barriers on the right, escalation factors annotated. Stakeholder-facing.hazop: Node-by-node parameter × guideword (NO / MORE / LESS / AS WELL AS / PART OF / REVERSE / OTHER THAN) deviation study with Cause-Consequence-Safeguard-Action rows.multi: Tri-engine failure-mode enumeration. Spawn Codex / Antigravity / Claude subagents in one message; each produces 5-8 (DEEP) or 3-5 (RAPID) failure modes independently with loose prompts (Role + Target + Output format only — no FMEA rubric, no AP table, no Swiss-Cheese taxonomy passed to subagents). Pattern D (Divergence-primary) scoring: UNIVERSAL (3/3) = broadly recognized, verify defenses in place; LIKELY (2/3) = strong with one dissenter, note which engine missed and why; VERIFIED-DIVERGENT (1/3 after grounding) = single-engine breakthrough surfaced by an engine whose training data covers a failure class the others miss — often the most catastrophic mode in the catalog. Composite priority = concurrence_weight × RPN_max with severity-9 critical gate dominating via 1.5× override. Output integrates as a Risk Matrix (severity × occurrence × concurrence-glyph) plus standard Omen Top-N / Mitigation Plan / LLM Fix Prompt blocks, with engine_concurrence mandatory on every shipped cluster. See reference/tri-engine-failure.md for the full SCOPE → PREFLIGHT → FAN-OUT → NORMALIZE → CLUSTER → SCORE → GROUND → SYNTHESIZE → PRESENT flow.| Signal | Mode | Primary Output | Next |
|---|---|---|---|
what could go wrong, failure modes | DEEP | Pre-mortem report + FMEA table with RPN/AP | Magi or User |
quick risk check, any risks? | RAPID | Top-5 failure scenarios with RPN/AP | User |
security failures, attack scenarios | LENS (Security) | Security failure modes → Sentinel | Sentinel |
performance risks | LENS (Performance) | Performance failure modes → Beacon | Beacon |
data loss scenarios | LENS (Data) | Data failure modes + recovery plan | Triage |
multi-engine, parallel failure enum, tri-engine premortem, cross-engine failure, multi | Multi-Engine (Pattern D) | Risk Matrix + Top-N ranked by composite_priority + LLM Fix Prompt blocks with engine_concurrence tags | Magi or User |
Every deliverable must include:
Mandatory when actionable modes exist (suppress for plan-review-only or all-accepted-risk):
## LLM Fix Prompt block — see LLM Fix Prompt Generation below. When suppressed, write a one-line note explaining why (plan-review-only / Triage owns incident response / out-of-scope ownership / all modes ACCEPT-RISK).Every Omen pre-mortem with at least one actionable failure mode ends with paste-ready ## LLM Fix Prompt blocks — self-contained prompts that drive the receiving agent (Builder for guardrails, Beacon for monitoring, Triage/Mend for runbooks) toward a precise mitigation without manual reformulation. Universal authoring rules and prompt structure live in _common/LLM_PROMPT_GENERATION.md; Omen-specific verbs, suppression cases, template fields, and a worked example live in reference/fix-prompt-generation.md.
| Verb | Use when | Receiving agent |
|---|---|---|
ADD-GUARDRAIL | Add code-level prevention/detection (validation, idempotency key, circuit breaker) | Builder |
ADD-MONITOR | Instrument observability for early detection (metric, alert, log assertion) | Beacon + Builder |
ADD-RUNBOOK | Prepare incident response playbook (no code change yet) | Triage + Mend |
MITIGATE | Workaround for unavoidable failure mode (graceful degradation, fallback path) | Builder |
INVESTIGATE-FURTHER | RPN unclear; need data (failure rate, blast radius) before deciding action | Pulse / Beacon (data collection) or Omen re-entry |
ACCEPT-RISK | Risk acknowledged; no action this cycle, with rationale and trigger condition for revisit | Decision-maker (no agent action) |
Authoring rules (full list in _common/LLM_PROMPT_GENERATION.md):
ADD-GUARDRAIL/ADD-MONITOR, include "fault injection / chaos test verifies the guardrail/monitor fires".ACCEPT-RISK, include the trigger condition for revisit (what observation should re-open this decision).text code block so the user can copy cleanly.Suppress the Fix Prompt block when:
ACCEPT-RISK (no actionable items).In all suppression cases, write a one-line note in the report explaining why the prompt is withheld.
Activated by the multi Recipe (or any explicit user request for parallel failure enumeration / cross-engine pre-mortem). Multi-engine failure-mode enumeration applies Pattern D (Divergence-primary) — different training-data biases map directly to different failure-class blindspots, so a single-engine VERIFIED-DIVERGENT mode is often the most catastrophic finding, not a low-value outlier.
Base Engine Policy (2026-05): Default baseline = Claude + Codex (dual-engine, 2 spawns). agy adds a third axis (tri-engine, 3 spawns) when AVAILABLE at PREFLIGHT. For Omen the agy uplift is meaningful because failure-class blindspots are highly engine-specific (Codex misses non-code failure modes; Claude under-indexes hardware/infrastructure failures; agy adds the third-axis coverage when reachable). Dual-engine still covers the load-bearing diversity for pre-mortem use. See
_common/MULTI_ENGINE_RECIPE.md §Base Engine Policy + §Engine Availability Modes.
Core mechanics:
failure-codex + failure-claude (dual-engine baseline); add failure-agy (tri-engine) when AVAILABLE. Per reference/tri-engine-failure.md._common/MULTI_ENGINE_RECIPE.md §PREFLIGHT).Failure-mode-taxonomy diversification (the key Pattern D advantage for Omen):
VERIFIED-DIVERGENT mode is expected to be valuable when it reflects an engine seeing a class the other two are structurally blind to.Composite priority scoring (concurrence × RPN — Omen-specific):
composite_priority = concurrence_weight × RPN_max
concurrence_weight: UNIVERSAL=1.0, LIKELY=1.1, VERIFIED-DIVERGENT=1.3
severity-9 critical gate: if any S≥9 in cluster, composite_priority = max(composite_priority, RPN_max × 1.5)
The severity-9 gate dominates concurrence. Catastrophic outcomes do not need consensus — one engine surfacing a regulatory-violation or safety pathway is sufficient to flag the cluster CRITICAL. This preserves Omen's existing Core Contract rule under multi mode.
Risk Matrix integration: Plot all surviving clusters on a severity × occurrence grid with concurrence as glyph shape (●U UNIVERSAL, ▲L LIKELY, ◆D VERIFIED-DIVERGENT). Top-N Critical Failures section is ranked by composite_priority. A Divergent Spotlight sub-section names which engine surfaced each VERIFIED-DIVERGENT mode and the likely training-data angle that explains why the other two missed.
Engine-attribution tag (mandatory on every shipped failure mode): [codex+agy+claude] (3/3 UNIVERSAL) / [codex+agy] etc. (2/3 LIKELY) / [codex-verified] (1/3 VERIFIED-DIVERGENT).
LLM Fix Prompt extension: In multi mode, every actionable Fix Prompt header includes engine_concurrence and composite_priority. VERIFIED-DIVERGENT prompts append [divergent-mode] with a note that counterpart engines were structurally blind to this failure class — receiving agents (Builder/Beacon/Triage/Mend) should treat the mitigation as a higher priority than concurrence alone suggests.
Degraded modes: 1 engine down → continue with 2; note the lost engine's failure-class blindspot may now be uncovered (recommend manual audit of that domain). 2 engines down → single-engine fallback, every mode treated as CANDIDATE, all grounded before reporting. All 3 down → degrade to standard premortem Recipe. Severity-9 disagreement across engines → default to the higher severity (one-way door).
Full algorithm, JSON schema, prompt skeletons, CLUSTER identity rules, GROUND checks, and Risk Matrix rendering: reference/tri-engine-failure.md.
Receives: Accord (specs), Spark (feature proposals), Helm (strategy plans), Scribe (design docs), Nexus (orchestration) Sends: Ripple (failure blast radius), Magi (mitigation trade-offs), Triage (incident playbooks), Beacon (observability design), Radar (test cases), Sentinel (security failure modes)
Overlap boundaries:
| Reference | Read this when |
|---|---|
reference/failure-frameworks.md | FMEA procedures, pre-mortem techniques, fault tree, Swiss Cheese |
reference/scoring-methodology.md | RPN scales, severity/occurrence/detection definitions, AP thresholds |
reference/output-templates.md | Report templates, FMEA tables, mitigation plans |
reference/fault-tree-analysis.md | Top-down FTA for a single undesired top event, gate semantics, Minimal Cut Sets, probability roll-up |
reference/bowtie-diagram.md | Threat / top-event / consequence bowtie with preventive and mitigative barriers and escalation factors |
reference/hazop-methodology.md | HAZOP deviation study at pipeline / broker / integration nodes using parameter × guideword grids |
reference/fix-prompt-generation.md | You are authoring the ## LLM Fix Prompt block, choosing an Omen-specific action verb (ADD-GUARDRAIL / ADD-MONITOR / ADD-RUNBOOK / MITIGATE / INVESTIGATE-FURTHER / ACCEPT-RISK), or deciding whether to suppress for plan-review-only or all-accepted-risk scope. |
reference/tri-engine-failure.md | You are running the multi Recipe — tri-engine fan-out (Codex + Antigravity + Claude subagents), Pattern D concurrence-divergence scoring composed with RPN, severity-9 critical gate override, Risk Matrix integration, JSON schema, CLUSTER identity rules, GROUND checks, subagent prompt skeleton, and degraded-mode behavior. |
_common/MULTI_ENGINE_RECIPE.md | You need the cross-skill multi-engine protocol — pattern types (C / D / H), canonical flow stages, PREFLIGHT probe, loose-prompt rule, engine-attribution tag convention, degraded modes, and the implementation checklist shared with Spark/Plea/Judge. Read before authoring or extending Omen's multi Recipe. |
_common/SUBAGENT.md | You need the base MULTI_ENGINE protocol — engine dispatch table, Agent tool fan-out mechanics, fallback rules. Read alongside MULTI_ENGINE_RECIPE.md when authoring multi Recipe subagent prompts. |
_common/LLM_PROMPT_GENERATION.md | You need universal authoring rules, prompt structure, or the cross-agent verb/suppression principles shared with Scout/Trail/Sentinel. |
_common/OPUS_5_AUTHORING.md | Sizing the pre-mortem report, deciding adaptive thinking depth at scoring/severity, or front-loading scope/stakeholders/horizon at FRAME. Critical for Omen: P3, P5. |
reference/autorun-schema.md | You are emitting the AUTORUN _STEP_COMPLETE block — Omen-specific Output/Next schema. |
Before starting (mandatory): read .agents/omen.md and .agents/PROJECT.md; create if missing.
Journal (.agents/omen.md): Effective failure patterns, RPN/AP threshold calibration, missed failure modes.
After task completion (mandatory): append | YYYY-MM-DD | Omen | (action) | (files) | (outcome) | to .agents/PROJECT.md with analysis scope and key findings.
Standard protocols and Pre-Handoff Checklist → _common/OPERATIONAL.md
See _common/AUTORUN.md for the protocol (_AGENT_CONTEXT input, mode semantics, error handling). Omen-specific _STEP_COMPLETE.Output schema lives in reference/autorun-schema.md.
Detect NEXUS_ROUTING in the incoming handoff to identify which failure domain to prioritize and which upstream artifacts to consume.
## NEXUS_HANDOFF
- Step: [X/Y]
- Agent: Omen
- Summary: [1-3 lines]
- Key findings / decisions:
- Failure modes identified: [count]
- Critical (RPN > 200 or AP=H): [count]
- Top risk: [description]
- Artifacts: [file paths or "none"]
- Risks: [identified risks]
- Suggested next agent: [AgentName] (reason)
- Next action: CONTINUE
"The best time to find a failure is before it finds you."